The Agent Skills Directory

[COMMAND_EXECUTION]: The file evals/files/report_service.py includes a subprocess.run call with shell=True using interpolated variables, which is a textbook command injection vulnerability. This is an intentional inclusion within the skill's evaluation dataset to test detection capabilities.
[DYNAMIC_EXECUTION]: The file evals/files/report_service.py uses pickle.loads() to deserialize user-supplied data, representing a critical remote code execution vector. This serves as a test case in the evals directory.
[CREDENTIALS_UNSAFE]: The files evals/files/ci-deploy.sh and evals/files/service-config.yaml contain hardcoded tokens and references to committed private keys. These are provided as example findings for auditing exercises.
[INDIRECT_PROMPT_INJECTION]: As a security auditing tool, this skill is designed to ingest and analyze untrusted source code. While the skill provides rigorous guidelines for triaging and dismissing false positives, the ingestion of external data constitutes an inherent indirect prompt injection surface.
[EXTERNAL_DOWNLOADS]: The evaluation file evals/files/api-package.json contains a postinstall script that fetches a remote artifact from an internal example CDN. This pattern is used in the skill's documentation to demonstrate supply-chain risks.

security-review